Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Valori: A Deterministic Memory Substrate for AI Systems

Varshith Gudur

TL;DR

This paper argues that bit-level determinism in AI memory is essential for replayability, auditability, and safety across hardware platforms. It proposes Valori, a no_std Rust kernel that replaces floating-point memory with fixed-point arithmetic ($Q16.16$) and treats memory as a deterministic state machine, enabling exact state serialization and cross-platform reproducibility. Valori additionally reworks indexing (e.g., $HNSW$) and precision contracts to ensure determinism without sacrificing semantic fidelity, achieving Recall@10 ≈ $99.8\%$ relative to a floating-point baseline and cross-platform snapshot equivalence. The work demonstrates practical feasibility with sub-millisecond latency and outlines a path toward auditable, consensus-ready AI systems, while noting current limitations and future extensions to higher-precision contracts and deterministic inference models.

Abstract

Modern AI systems rely on vector embeddings stored and searched using floating-point arithmetic. While effective for approximate similarity search, this design introduces fundamental non-determinism: identical models, inputs, and code can produce different memory states and retrieval results across hardware architectures (e.g., x86 vs. ARM). This prevents replayability and safe deployment, leading to silent data divergence that prevents post-hoc verification and compromises audit trails in regulated sectors. We present Valori, a deterministic AI memory substrate that replaces floating-point memory operations with fixed-point arithmetic (Q16.16) and models memory as a replayable state machine. Valori guarantees bit-identical memory states, snapshots, and search results across platforms. We demonstrate that non-determinism arises before indexing or retrieval and show how Valori enforces determinism at the memory boundary. Our results suggest that deterministic memory is a necessary primitive for trustworthy AI systems. The reference implementation is open-source and available at https://github.com/varshith-Git/Valori-Kernel (archived at https://zenodo.org/records/18022660).

Valori: A Deterministic Memory Substrate for AI Systems

TL;DR

This paper argues that bit-level determinism in AI memory is essential for replayability, auditability, and safety across hardware platforms. It proposes Valori, a no_std Rust kernel that replaces floating-point memory with fixed-point arithmetic () and treats memory as a deterministic state machine, enabling exact state serialization and cross-platform reproducibility. Valori additionally reworks indexing (e.g., ) and precision contracts to ensure determinism without sacrificing semantic fidelity, achieving Recall@10 ≈ relative to a floating-point baseline and cross-platform snapshot equivalence. The work demonstrates practical feasibility with sub-millisecond latency and outlines a path toward auditable, consensus-ready AI systems, while noting current limitations and future extensions to higher-precision contracts and deterministic inference models.

Abstract

Modern AI systems rely on vector embeddings stored and searched using floating-point arithmetic. While effective for approximate similarity search, this design introduces fundamental non-determinism: identical models, inputs, and code can produce different memory states and retrieval results across hardware architectures (e.g., x86 vs. ARM). This prevents replayability and safe deployment, leading to silent data divergence that prevents post-hoc verification and compromises audit trails in regulated sectors. We present Valori, a deterministic AI memory substrate that replaces floating-point memory operations with fixed-point arithmetic (Q16.16) and models memory as a replayable state machine. Valori guarantees bit-identical memory states, snapshots, and search results across platforms. We demonstrate that non-determinism arises before indexing or retrieval and show how Valori enforces determinism at the memory boundary. Our results suggest that deterministic memory is a necessary primitive for trustworthy AI systems. The reference implementation is open-source and available at https://github.com/varshith-Git/Valori-Kernel (archived at https://zenodo.org/records/18022660).
Paper Structure (23 sections, 2 equations, 1 figure, 3 tables)

This paper contains 23 sections, 2 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Background & Motivation
  3. Floating-Point Non-Determinism
  4. The Embedding Pipeline
  5. Problem Definition
  6. Memory State and Replayability
  7. Empirical Evidence of Non-Determinism
  8. Reproduction Setup
  9. Results
  10. Valori Design
  11. Fixed-Point Arithmetic (Q16.16)
  12. Memory as a State Machine
  13. Architecture Separation
  14. Precision as a Configurable Memory Contract
  15. Indexing and Determinism
  16. ...and 8 more sections

Figures (1)

  • Figure 1: Valori Architecture: The Kernel operates as a pure deterministic core, wrapped by interface layers (Python FFI or Node HTTP) handling I/O.